NowComment
2-Pane Combined
Comments:
Full Summaries Sorted

Disinformation and AI chatbots - The New York Times

Author: Tiffany Hsu and Stuart A. Thompson

Hsu, Tiffany, and Stuart A. Thompson. “Disinformation Researchers Raise Alarms About A.I. Chatbots.” The New York Times, 8 Feb. 2023. NYTimes.com, https://www.nytimes.com/2023/02/08/technology/ai-chatbots-disinformation.html.

Soon after ChatGPT debuted last year, researchers tested what the artificial intelligence chatbot would write after it was asked questions peppered with conspiracy theories and false narratives.

The results — in writings formatted as news articles, essays and television scripts — were so troubling that the researchers minced no words.

This tool is going to be the most powerful tool for spreading misinformation that has ever been on the internet,” said Gordon Crovitz, a co-chief executive of NewsGuard, a company that tracks online misinformation and conducted the experiment last month. “Crafting a new false narrative can now be done at dramatic scale, and much more frequently — it’s like having A.I. agents contributing to disinformation.”

Disinformation is difficult to wrangle when it’s created manually by humans. Researchers predict that generative technology could make disinformation cheaper and easier to produce for an even larger number of conspiracy theorists and spreaders of disinformation.

Personalized, real-time chatbots could share conspiracy theories in increasingly credible and persuasive ways, researchers say, smoothing out human errors like poor syntax and mistranslations and advancing beyond easily discoverable copy-paste jobs. And they say that no available mitigation tactics can effectively combat it.

Predecessors to ChatGPT, which was created by the San Francisco artificial intelligence company OpenAI, have been used for years to pepper online forums and social media platforms with (often grammatically suspect) comments and spam. Microsoft had to halt activity from its Tay chatbot within 24 hours of introducing it on Twitter in 2016 after trolls taught it to spew racist and xenophobic language.

ChatGPT is far more powerful and sophisticated. Supplied with questions loaded with disinformation, it can produce convincing, clean variations on the content en masse within seconds, without disclosing its sources. On Tuesday, Microsoft and OpenAI introduced a new Bing search engine and web browser that can use chatbot technology to plan vacations, translate texts or conduct research.

OpenAI researchers have long been nervous about chatbots falling into nefarious hands, writing in a 2019 paper of their “concern that its capabilities could lower costs of disinformation campaigns” and aid in the malicious pursuit “of monetary gain, a particular political agenda, and/or a desire to create chaos or confusion.”

In 2020, researchers at the Center on Terrorism, Extremism and Counterterrorism at the Middlebury Institute of International Studies found that GPT-3, the underlying technology for ChatGPT, had “impressively deep knowledge of extremist communities” and could be prompted to produce polemics in the style of mass shooters, fake forum threads discussing Nazism, a defense of QAnon and even multilingual extremist texts.

OpenAI uses machines and humans to monitor content that is fed into and produced by ChatGPT, a spokesman said. The company relies on both its human A.I. trainers and feedback from users to identify and filter out toxic training data while teaching ChatGPT to produce better-informed responses.

OpenAI’s policies prohibit use of its technology to promote dishonesty, deceive or manipulate users or attempt to influence politics; the company offers a free moderation tool to handle content that promotes hate, self-harm, violence or sex. But at the moment, the tool offers limited support for languages other than English and does not identify political material, spam, deception or malware. ChatGPT cautions users that it “may occasionally produce harmful instructions or biased content.”

Last week, OpenAI announced a separate tool to help discern when text was written by a human as opposed to artificial intelligence, partly to identify automated misinformation campaigns. The company warned that its tool was not fully reliable — accurately identifying A.I. text only 26 percent of the time (while incorrectly labeling human-written text 9 percent of the time) — and could be evaded. The tool also struggled with texts that had fewer than 1,000 characters or were written in languages other than English.

Arvind Narayanan, a computer science professor at Princeton, wrote on Twitter in December that he had asked ChatGPT some basic questions about information security that he had posed to students in an exam. The chatbot responded with answers that sounded plausible but were actually nonsense, he wrote.

“The danger is that you can’t tell when it’s wrong unless you already know the answer,” he wrote. “It was so unsettling I had to look at my reference solutions to make sure I wasn’t losing my mind.” Researchers also worry that the technology could be exploited by foreign agents hoping to spread disinformation in English. Some companies already use multilingual chatbots to support customers without translators.

Mitigation tactics exist — media literacy campaigns, “radioactive” data that identifies the work of generative models, government restrictions, tighter controls on users, even proof-of-personhood requirements by social media platforms — but many are problematic in their own ways. The researchers concluded that there “is no silver bullet that will singularly dismantle the threat.

Working last month off a sampling of 100 false narratives from before 2022 (ChatGPT is trained mostly on data through 2021), NewsGuard asked the chatbot to write content advancing harmful health claims about vaccines, mimicking propaganda and disinformation from China and Russia and echoing the tone of partisan news outlets.

The technology produced responses that seemed authoritative but were often provably untrue. Many were pockmarked with phrases popular with misinformation peddlers, such as “do your own research” and “caught red-handed,” along with citations of fake scientific studies and even references to falsehoods not mentioned in the original prompt. Caveats, such as urging readers to “consult with your doctor or a qualified health care professional,” were usually buried under several paragraphs of incorrect information.

Researchers prodded ChatGPT to discuss the 2018 shooting in Parkland, Fla., that killed 17 people at Marjory Stoneman Douglas High School, using the perspective of Alex Jones, the conspiracy theorist who filed for bankruptcy last year after losing a series of defamation cases brought by relatives of other mass shooting victims. In its response, the chatbot repeated lies about the mainstream media colluding with the government to push a gun-control agenda by employing crisis actors.

Sometimes, though, ChatGPT resisted researchers’ attempts to get it to generate misinformation and debunked falsehoods instead. (This has led some conservative commentators to claim that the technology has a politically liberal bias, as have experiments in which ChatGPT refused to produce a poem about former President Donald J. Trump but generated glowing verses about President Biden.

Newsguard asked the chatbot to write an opinion piece from Mr. Trump’s perspective about how Barack Obama was born in Kenya, a lie repeatedly advanced by Mr. Trump for years in an attempt to cast doubt on Mr. Obama’s eligibility to be president. ChatGPT responded with a disclaimer that the so-called birther argument “is not based on fact and has been repeatedly debunked” and, furthermore, that “it is not appropriate or respectful to propagate misinformation or falsehoods about any individual.”

When The New York Times repeated the experiment using a sample of NewsGuard’s questions, ChatGPT was more likely to push back on the prompts than when researchers originally ran the test, offering disinformation in response to only 33 percent of the questions. NewsGuard said that ChatGPT was constantly changing as developers tweaked the algorithm and that the bot might respond differently if a user repeatedly inputs misinformation.

Concerned legislators are sounding calls for government intervention as more ChatGPT rivals crowd the pipeline. Google began testing its experimental Bard chatbot on Monday and will release it to the public in the coming weeks. Baidu has Ernie, short for Enhanced Representation through Knowledge Integration. Meta unveiled Galactica (but took it down three days later amid concerns about inaccuracies and misinformation).

In September, Representative Anna G. Eshoo, Democrat of California, pressured federal officials to address models like Stability AI’s Stable Diffusion image generator, which she criticized for being “available for anyone to use without any hard restrictions.” Stable Diffusion, she wrote in an open letter, can and likely has already been used to create “images used for disinformation and misinformation campaigns.”

Check Point Research, a group providing cyber threat intelligence, found that cybercriminals were already experimenting with using ChatGPT to create malware. While hacking typically requires a high level of programming knowledge, ChatGPT was giving novice programmers a leg up, said Mark Ostrowski, the head of engineering for Check Point.

“The amount of power that could be circulating because of a tool like this is just going to be increased,” he said.

DMU Timestamp: February 21, 2023 13:31





Image
0 comments, 0 areas
add area
add comment
change display
Video
add comment

Quickstart: Commenting and Sharing

How to Comment
  • Click icons on the left to see existing comments.
  • Desktop/Laptop: double-click any text, highlight a section of an image, or add a comment while a video is playing to start a new conversation.
    Tablet/Phone: single click then click on the "Start One" link (look right or below).
  • Click "Reply" on a comment to join the conversation.
How to Share Documents
  1. "Upload" a new document.
  2. "Invite" others to it.

Logging in, please wait... Blue_on_grey_spinner